Memory Scores
The Memory Platform uses a multi-dimensional scoring system to determine the relevance, quality, and longevity of each piece of knowledge.
Core Scores
Salience (0.0 - 1.0)
Salience represents how important or "front-of-mind" a memory is.
- Initial Value: Determined by the extractor based on the perceived importance of the information.
- Access Boost: Each time a memory is retrieved for context, its salience is increased.
- Time Decay: Salience gradually decreases over time if the memory is not accessed.
- Usage: Primary factor in ranking memories for AI context.
Confidence (0.0 - 1.0)
Confidence represents the system's certainty that a memory is accurate.
- Source Weight: Memories from explicit user statements have higher confidence than those inferred from patterns.
- Verification: Confidence can increase if the same fact is mentioned multiple times or verified by a user.
- Usage: Filtering out low-quality or uncertain information.
Stability (0.0 - 1.0)
Stability represents how resistant a memory is to change or invalidation.
- High Stability: Birthdays, founding dates, physical addresses.
- Low Stability: Current project status, temporary moods, recent opinions.
- Usage: Determining the decay rate (Stable memories decay much slower).
The Decay Algorithm
The system applies exponential decay to the Salience score to simulate how "remembrance" naturally fades.
Exponential Decay Formula
The salience at time t is calculated as:
salience(t) = salience(0) * e^(-λ * days)
Where:
λ(Lambda) is the decay rate, which depends on theMemoryKindandStability.daysis the time elapsed since the last access.
Default Decay Rates (λ)
| Memory Kind | Decay Rate (λ) | Description |
|---|---|---|
| Fact | 0.01 | Very slow (years) |
| Preference | 0.05 | Moderate (months) |
| Insight | 0.10 | Fast (weeks) |
| Summary | 0.15 | Very fast (days) |
Composite Score
When retrieving memories, the system calculates a Composite Relevance Score to rank the results:
Relevance = (Semantic Similarity * w1) + (Salience * w2) + (Confidence * w3)
This ensures that the AI receives context that is both semantically relevant to the current query and "salient" in the system's current understanding.